LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
NotíciasTópico · 模型压缩与效率Material2 min de leituraEditorial DisconfirmAI

AI startup Inception launches Mercury 3, a diffusion-based LLM claiming ~1,000 tokens per second generation speed

Inception releases Mercury 3, a flagship diffusion large language model that claims to significantly outperform traditional autoregressive models in inference speed.

2026-10-20 05:49 GMT+8

US AI startup Inception launched Mercury 3 on October 19, 2026. The new flagship large language model uses diffusion technology rather than the traditional autoregressive architecture, generating and refining multiple tokens in parallel to drastically improve inference speed. Inception CEO Stefano Ermon told The Deep View that Mercury 3 can generate approximately 1,000 tokens per second, making it 4x to 8x faster than Anthropic's Haiku 5.5, OpenAI's GPT-6 Luna, and Zhipu AI's GLM-5.3 Flash.

Regarding pricing, Mercury 3 is currently in its launch promotion period, costing $0.10 per million input tokens and $0.35 per million output tokens. Ermon stated that this price is 50% off the eventual standard rate, and the company is working to make the promotional price permanent in the coming weeks. While the pricing is competitive, Inception emphasizes that Mercury 3's primary advantage is speed rather than low cost, aiming to match or exceed the performance of current mainstream low-to-mid-tier models.

It is important to note that the claims regarding speed multipliers, performance parity, and being the "most capable diffusion-based language model" are based on Inception's own testing and executive statements; there are currently no public third-party independent benchmarks verifying these results. Additionally, Inception has not disclosed the exact total parameter count of Mercury 3, only stating it is 40 times larger than Mercury 1 released in February 2025, which makes it difficult for outsiders to precisely evaluate its energy efficiency under equivalent compute budgets.

Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

Fontes

  • thedeepview.com↗
Tópico

模型压缩与效率

AI行业正通过架构创新与极致压缩挑战传统大模型的算力瓶颈。美国初创公司Inception Labs于2026年10月20日发布旗舰扩散式大语言模型Mercury 3,宣称其生成速度达每秒1000个token,推理成本显著低于Anthropi…

2 etapas2 peçasAtualizado 2026-10-21 04:52

Como chegamos aqui

  1. 2026-10-20Inception Labs发布扩散式大语言模型Mercury 3,CEO Stefano Ermon称其速度达1000 tokens/秒,价格比竞品低35%,并声称在代理基准测试中表现优于Haiku 5.5等模型。thedeepview.com ↗
  2. 2026-10-20PrismML发布Bonsai 2 27B,基于Qwen3.8 27B进行三元权重压缩,体积降至5.9GB,基准测试得分达原版98%,旨在支持本地设备运行。techorange.com ↗

O que ainda está indefinido

  • Mercury 3的扩散架构在处理长上下文逻辑推理时,是否能持续保持与自回归模型相当的质量?
  • Bonsai 2等压缩模型在实际用户场景中的体验损失(如幻觉率增加)是否真的如基准测试所示微乎其微?
  • 随着参数规模扩大,压缩模型能否真正实现100%性能保留,还是存在不可逾越的信息熵极限?

Também nesta thread

  • AI Startup PrismML Releases Bonsai 2: Compresses 27B Model to 5.9GB with 98% Self-Tested Performance Retention2026-10-20
Ler o tópico completo →
Continuar pesquisando →